abdulmannaan1/ai-masters-advisor-france

🎓 AI Masters Advisor – France

This Space hosts a Retrieval-Augmented Generation (RAG) system designed to answer student questions about applying to Master’s programs in France.

The assistant supports English and French queries and is grounded strictly on official academic documents, including:

• —Campus France "Études en France" guide
• —Campus France Master's admission documentation
• —Université Paris-Saclay AI Master's catalogue
• —Grenoble AI4OneHealth Master's guide
• —HEC MSc Data Science program brochure

All answers are generated using a retrieval + grounding pipeline to minimize hallucinations and ensure factual accuracy.

🔍 System Features

✅ Multilingual question answering (English & French) ✅ FAISS vector retrieval over official university PDFs ✅ Phi-3 LLM response synthesis ✅ Faithfulness + Recall@k benchmarking ✅ Fully reproducible evaluation pipeline

📊 Evaluation Results

Benchmarking was conducted on bilingual admissions QA sets:

The results demonstrate strong cross-lingual grounding performance with higher top-rank retrieval precision observed for French queries.

🔗 Project Links

• —✅ Source code & experiments: https://github.com/abdulmannaan502/french-masters-rag-advisor

• —✅ Reproducible evaluation notebook (Kaggle): (Link will be added after notebook publication)

⚙️ Architecture Overview

Pipeline Flow:

PDF documents → Chunking → Embedding → FAISS Vector Search → Top-K Grounded Context → Phi-3 Generation → Verified Answer

All outputs are constrained to retrieved document context to maintain factual reliability.

🎯 Use Case

This assistant is intended as:

• —A proof-of-concept admissions advisor chatbot
• —A research artifact for multilingual NLP evaluation
• —A portfolio system for graduate AI program applications

📄 Research

A technical evaluation of this system is documented in a short research preprint:

_“Multilingual Evaluation of a Retrieval-Augmented Generation System for Admissions Question Answering”_

(arXiv submission pending)

🛠 Maintenance

This project is actively maintained and expanded for:

• —Larger bilingual datasets
• —Retrieval reranking experiments
• —Response faithfulness auditing